摘要
针对高速铁路探伤巡检过程中普遍存在的训练样本不平衡和打标签复杂等问题,提出一种改进的Skip-GAN算法,对巡检图像中存在的多类病害进行无监督检测。对高速铁路巡检图像进行预处理,包括轨道板分割和数据增强,减少排水沟和明暗环境对网络重构图像的影响;对Skip-GAN结构进行改进,包括增加注意力机制模块、修改判别器为双自动编码器结构等;改进卷积神经网络的损失函数,增加网络对图像的重构能力;将高速铁路探伤巡检中正常图像作为训练样本输入模型进行训练,实现端到端的巡检图像多种类病害的检测。实验结果表明:提出的检测模型在探伤图像病害样本数量少时,对钢轨表面伤损、扣件缺少、轨道板异物三种类型的异常检测取得良好的检测结果,模型的识别精确度、F1、AUC分别达到0868、0821、0842。
In response to the problems of unbalanced training samples and complex labeling in the process of high-speed rail flaw detection,this paper proposed an improved Skip-GAN algorithm to perform unsupervised detection of multiple types of defects in inspection images.First,the high-speed railway inspection images were preprocessed,including track slab segmentation and data enhancement,to reduce the influence of drainage ditches and light and dark environments on network reconstructed images.Second,improvements were made to the Skip-GAN structure,including increasing an at-tention mechanism module and modifying the discriminator to a dual auto-encoder structure.Third,the loss function of the convolutional neural network was improved to increase the network ability to reconstruct images.Finally,the normal images in the high-speed rail flaw detection inspection were used as the training sample input model for training,reali-zing end-to-end detection of various defects in inspection images.The experimental results show that the detection model proposed in this paper achieves good detection results for three types of abnormal detection of rail surface damage,miss-ing fasteners,and foreign matters in the track slab in the case of small number of defect samples with the precision,F1,and AUC of the model reaching 0868,0821,and 0842,respectively.
作者
何庆
刘震
王启航
张岷
王平
HE Qing;LIU Zhen;WANG Qihang;ZHANG Min;WANG Ping(School of Civil Engineering,Southwest Jiaotong University,Chengdu 610031,China;Key Laboratory of High-speed Railway Engineering,Ministry of Education,Southwest Jiaotong University,Chengdu 610031,China;China Railway First Survey and Design Institute Group Co.,Ltd.,Xi’an 710043,China)
出处
《铁道学报》
EI
CAS
CSCD
北大核心
2024年第9期121-128,共8页
Journal of the China Railway Society
基金
国家自然科学基金(51878576,U1934214)
陕西省重点研发计划(2021ZDLGY02-03)。
关键词
高速铁路
巡检图像
异常检测
GAN
无监督学习
深度学习
high-speed railway
patrol inspection image
anomaly detection
GAN
unsupervised learning
deep learning
